Google DeepMind’s DiffusionGemma shows that a text diffusion model does not have to be trained from scratch.
The Decoder reports that DeepMind converted Gemma-4-26B-A4B into a diffusion-style language model using less than 10% of the original training budget. Instead of generating text one token at a time, DiffusionGemma refines blocks of 256 tokens in parallel, similar to how image diffusion systems gradually remove noise.
On an Nvidia H100 accelerator, the model reportedly reaches about 1,500 tokens per second. That makes the approach interesting for applications where latency and throughput matter, especially if parallel generation can be made reliable enough for production use.
The tradeoff is quality. The article notes that DiffusionGemma still trails the strongest autoregressive language models. Its importance is more about the engineering path: existing models may be reusable as starting points for faster generation methods, reducing the cost of experimentation.